- -- . ...- .— - ...., ..— The dynamics of mass interaction Steve l~ttaker, Loren Terveen, W~ Hti, Lynn Cherny An Labs-Research 180 ParkAve FlorhamPark NJ 07932 USA +1 9733608339 {stevew,terveeW_,cherny} @esearch.att.com over 17,000 newsgroups, with approximately 3 tion users worldwide [7]. ~s groti has been achieved without any cenmfied organisation or governing body [9]. me ubiquity of Usenet and the fact that it suppoti conversations between hundreds or even thousands of pticipants, provides the opportunity to study what we term mass interaction. However, we currentiy lack basic data about Usenet interactions. me current paper analyses over 2.15 Won messages produced by 659,450 people in 500 representative newsgroups co fleeted over 6 months. We provide descriptive dab about newsgroup demographics, communication strategies and interactivity. We hen derive predictions from tie common pound model of communication to test predictions about how these parametersinteract ABSTRACT Usenet may be regarded as the worl$s largest conversatioti appficatioq with over 17,000 newsgronps and 3 fion users. Despite its ubiquity and poptiari~, however, we know Etie about the natore of the interactions it suppo~. ~s empirical paper investigates mass interaction in UseneL We anrdyse over 2.15 Mon messages from 659,450 posters, co~ected from 500 newsgoups over 6 months. We fit characterise mass interactio~ presenting basic data about demographics, conversational strategies and interactivity- Using predictions tiom the common ground [3] model of interactio~ we next conduct causal modetig to determine relations bemeen demographics, conversational strategim and interactiv~. We find evidence for moderate conversatiomd threading, but hge participation ine@ties in Usene4 with a d minority of participants posting a large proportion of messages. Contrary to the common gromd model and ‘Etiquette” guidefies [8,10] we ho find that “cross-posting” to extemd newsgroups is hig~y fiequenk Our predictions about the eEects of demographics on conversatioti strategy were largely cotie~ but we found disconfirming evidence about the relations befi’een conversational strategy and interactivity. Contrary to our expectations, boti cross-posting and short messages promote interactivity. We conclude that in order to e~lain mass interaction, the common ground model must be mowed to ticorporate notions of weak ties [5,q and communication overload [11,18]. Previous research on Usenet has tended to carry out sma~ sde qtitative studies of specific newsgroups, their ctitie and tieir conversation [1,17]. me these studies have drawn attention to impotit phenomena, their specfic focus means tiey cannot address general questions tit are central to mass interaction such as the levels of communication betieen ~erent newsgroups. ~ey also cannot easly examine the effects on interaction of different demographic variables or communication strategies, in order to test specXc communication models. me current study attemptsto redress the balance. We f~st present basic information from the 500 newsgroups about mass interaction addressing the fo~owing questions. What are tie dernogaphics of a typical newsgroup: how many,. people contribute and how often do they do so? Is participation roufly equal or are groups dominated by a few verbose individurds?A second set of questions concern conversational strate~-es. Do participants restrict their interactions to the current newsgroup or do they broadcast them widely to mtitiple groups (a phenomenon known as cross-posting)? How long is a typical message? Do most newsgroups have FAQs ~Cfists of tiequenfly asked questions”) and how often do they post them? Finafly we can ask questions about interactivi~ how deep is a typical conversational threa~ and how often are attempts to initiateconversation successti? Ke~ords hlass interactio~ Usene< conversatio~ newsgroups, common groun~ moderatio~ FAQS, netiquette, empirid. lNTRODUC~ON Usenet maybe regarded as the worl&s largest and fmtest growing conversationrd appficatiom k 1988 there were fewer than 500 news~oups. bent estimatesvary, but at the time of our dab co~ection in Dec. 1996, there were Ptission 10de distal or hard copies ofall or pti ofthis ~vorkfor personal or classroom use is granted without f= protided that copies Zre notmade or dis~%uted for profit or commercial advantage and that copies bear this notice and the fill citation on the fimt paga To copy othm~ise to republish, to post on sen’em or to redistribute to Iis& Tquires prior sptific permission and~ora fee CSCW 98 Seatie Wastinson USA ‘Cop3ti@tACLI 199s 1-581i34094/98/11-..Soooo I --— 37 -. *. . ,. ,, —— —.— . . . _.. We *O tested how weu the common pound communication theory [3] expw mass interaction The sc~e of mass interactio~ and the huge numbers of messages and participants, gives rise to WO novel communication problems. The &t of these concern how participants estabkh common ground in mass interaction Co=on ground is a key principle of fice-to-face conversation and refers to tie fact that participants must estabhh a degree of muti knowledge for their conversational contributions to be understood [3]. Face-b face conversations genetiy take place between dyads or smti groups, however. So how can conunon ground be estabtihed when there is a huge set of conversationrd participants with potenMy diverse perspectives? There are rdso issues concerning the stabtity of the newsgroup pop&tioL h some newsgroups, there is a core set of participantswho repeatedy converse, flowing participants to become ftiar with one another. This ftiarity supports conunon groun~ but how do conversation fae in newsgroups where the participants are constantly changing? Some newsgroups employ a strategy of moderation to addressthe problem of common ground h a moderated group, W interactions are fltered by a d set of moderators who are knowledgeable about the gords and history of the newsgroup. My messages that are relevant to the newsgroup’s goti are dewed to appear. Our fit set of hypotheses concern how the three demographic tictors of newsgroup ske, the familian-~ of participants, and the effects of moderation affect common ground We *O investigate how these same factors &ect interactivip. One measure of conversationrd interactivity is the extent to which a given conversationrd contribution depends on prior context [3,14]. By this defitio~ deeply threaded conversations are indicators of interactivity. We therefore looked at how demographic fictors tiect threading. The second major connnunication problem arises directiy horn the ~cd~ of estabtihing common ground bck of common ground may mean that participants are inundated with postings that are redundant or irrelevant For example, new participants who are ignorant of a newsgroup’s go~ or conversation history, may post questions that have been discussed before, or are largely fielevant to the group’s interests.We therefore investigated a second set of factors concerning people’s conversational strategim in addressing redundancy and irrelevance. We tested the effects of three conversatioti smtegies on interactivity: (a) FAQ production; @) long messages; (c) low levek of cross-posting. ~ese strategies are ~ derived from the notion of common ground FAQs exernp~ common ground by ~ing prior discussions and providing tiormation about newsgroup ctitiq cross-posting can be seen as a f~ure of common ground representing participants’ need to go outside the newsgroup for conversation long messages can be taken as evidence for substantive discussions and hence estab~ed common . a.- ..- ‘~>1-‘-. ground. The common ground model would therefore predict that FAQ productio~ decreased cross-posting and greater message length shodd au increase titeractivity. A second set of partia~y overlapping predictions about the effects of conversational strategy on interactivity, can be derived from the Netiquette guidekes (news.announce.newusers and news.answers). These area set of prescriptive guideties about effective communication strategies [S,10]. They tell users to be succinct ~avoid long postings”), avoid redundantor repeat postings ~read the FAQ’), and to avoid widespread posting of messages of ody marginal relevance ~don’t Cross-posfl). With the exception of succinctness, tie Netiquette guidekes and common ground model make identicrd predictions about the effects of conversational strategieson interactivity. To summarise, the paper first presents basic descriptive statistics for newsgroups: (a) demographics - stie, ffiarity and moderatio~ @) conversation strategies FAQ productio~ message len~ and cross-posting; (c) interactivity - the extent of conversational threading. We then use predictions derived from the common ground model to test three further questions: (a) how the three demographic variables affect conversational strategy; @) how demographic variables affects interactivity; (c) how conversational strategy affects interactivity. We conclude with a general characterisation of mass interaction and a discussion of how we~ the common ground model apphes to mass interaction. USENET ORGANISATION, INTERACTION DISTRIBUTION AND There are several Ievek of s~cture in Usenet. CoUections of messages are clustered into newsgroups, and newsgroups themelves are organised into hierarchies. Each hierarchy is intended to address different conversational topics. There are over a hundred different hierarchiesbut the majority of newsgroups belong to one of eight main hierarchies ~’the big eight”). These eight are: alt for alternative topics; comp for computer issues; humanitia; mkc for misce~aneous discussions; nws for discussions about Usene~ rec for recreational topics; soc for social issues; and talk for general conversations. Each hierarchy is then subdivided into more specfic subhierarchies. The name of each newsgroup begins with the relevant main hierarchy, and tem of increasing specificity are added to this. Typical newsgroup names are rec.music.dylan and rec.music.beatles, where both newsgroups are part of the music subhierarchy within rec. Twelve percent of newsgroups are moderated, and moderation is reflected in the newsgroup name: soc.russian.ctiture.moderated. Newsgroup distribution also varies from site to site. Newsgroups are distributedthrough an infod network of servers, with the specific set of .— —. —___ ..——- . . groups. These messages contain computer programs or images and we were ody interested in textual messages, since we are studying conversation. For the same reaso~ we excluded n~vs, *bk*, and ~obs * groups, because these groups are places for generrd announcements and advertisements, not conversation. We excluded regional groups, as our focus was on groups with global distriiutiom Fiagure1: Interaction md threading in r~.mwic.d>~Ian A third level of structurewithin newsgroups is achieved by a mechanism that trach conversationrdthreads (see Figure 1). A conversation begins with an initiating post about a nem7 issue or question ~Re Robert Johnson and Don Harnmon~, posted by HW61) and any messages responding to the new posting ti be Med to the message to which they respon& This is a conversationrdthrea~ i.e. a conversation consists of an tiitiating post along with various responses (if the posting eficits responses). There are &o mtitiple Ievek of threading, so responding to an initialpost is different from responding to a response to the initial posL For e-le, Figure 1 shows that the fired message in the ‘Robert Johnson” thread (from ~ routier) was acwy a response to tie second posting in fiat thread(Horn Sadiejane), as revealed by the be joining them hfost newsreading software organises messages according to threads, so users can view conversations in terms of the initiating posL They can rdso view the tiead depth of a given message, ie. whether it is an initiating post (depth of zero) a response an initiating post (depth of 1), and so OE FMy, cross-posting a message to mtitiple newsgroups is done by entering the names of additionrd newsgroups in the nen~sgroups field wMe composing a message. DATA COLLEC~ON AND SMPLING We sampled from 26 Merent to~level categories, inclutig seven of the ‘%ig eight” categories. We excluded 17ariousnewsgroups for Merent reasons. We first excluded altgroups. hlmy server administratorshave a poficy of not distributing alt groups because of the tone and subject matter of many alt discussions. This *es alt groups less ubiquitous in terms of their distriiutiom Alt groups &o ~er horn N other soups in their approval and creation process. T$7e*O excluded hmzanitim groups because there are ordy a few groups fiess b 10 worldwide thatwe have been able to find). We did not coflect data for any binaries — .,.— — This left us considering groups from the categories comp, mist, ret, sci, SOC, and talk. From our newsfee~ we selected a stratied random sample of 500 newsgroups from these groups in proportion to their global occurrence, where the proportions were derived from a master list of 17,112 currendy active newsgroups dodoaded from uunet (ftp.uu.net). We selected ody “active” groups, which we defined as those groups for which tiere were at least 1S0 messages over 6 months, equivalent to one message per day. A til selection criterion was moderation the groups selected were representative of the overa~ level of moderation in UseneL We coflected header information about each message in each of those newsgroups over a six month perio~ from Jtiy to December 1996. For each message we extracted various types of tiormation including: the ed addressof the message posteq the date that the message was poste~ the subject he of the message; message length (number of hes in the body of the message); and thread depth. h cases where messages were cross-posted we also gathered information about the newsgroups that the message was posted to. We first present descriptive data about newsgroup demographics, conversation strate~”m and interactive@. We then outie spec%c hypotheses about the relations beween these factors, and present a causal model testing these hypotheses. DESCRIPTIVE RESULTS Demogmpti= Table 1 shows demographic statistics for the 500 newsgroups. For each newsgroup, we calcdated the total number of messages that people posted over the 6 month interv~ and the number of different people who posted to that newsgroup. Overti levek of message traffic were hi~ with each group on average receiving 4299 messages (an average of 24 messages per newsgroup per day). h additio~ each newsgroup attracted contributions from an average of 1319 Merent posters. Together these statistics provide evidence for mass interaction. To calcdate thefamiliarip of posters in a newsgroup, we used &e criterion of repeat posting, i.e. whether a person posts more than once to a newsgroup. Our datashow thatin gened a si@cant proportion of users are titic 27% of messages are from “singleton posters” who ody contribute to a newsgroup once. This raised the issue of ——- ● .—.. — -.—. — whe&er newsgroups have participation ineqtity. Some participants must be contributing large numbers of messages, given (a) the high proportion of singIeton poster$ and @) that the mean level of contiiution is 3.1 messages/poster. Figure 2 shows hi~y uneqti Ievek of pardcipation in each newsgroup. The right hand side of the ~=ph shows that the majority of people post ody a few times, wMe the left hand side revefi that there are a few people who post a large number of times. A fid statistic bearing this out is tiat a tiny percentage (2.9%) of posters ti each newsgroup account for an average 25~0 of the toti posts. DATA COLLE=D Number pardcipants I Demographics I I 0/0messages horn repeatposters Number messages I hlessage length I Conversatioti ‘/0 messages tit were cross- hlssage byPosters (rink@ ● a 1319 73 Figure 2: Levels newsgroup 4299 44 Thread depti individual participation within Interactivity 34 Fina~y we looked at interactivi~. The average thread depth is 1.S messages, suggesting atypical exchange in which the average message refers to approximately two other messages. A substantialproportion of messages (33Yo) had two or more threads, indicating frequent extended conversations. We ako calcdated a complementary measure of interactivity, the number of initiating posts (those with zero references and thread depth of one). A high proportion of initiating messages would indicate repeated fadures to start conversations, or a prevalence of conversational dead-ends. We found that initiating messages are hi~y prevalen$ accounting for over more than 40% of messages. This suggests a view of Usenet in which it is hard to starta conversation. Once a conversation starts, however, then it seems to atiact mtitiple contributions. + 0.4 bteractivity of 1 1.s Table 1: D~criptive statisticsfor the 500 newsgoups over the 6 nzo?zthsof the stud~~ Conversational Shtegies lt7e found cross-posting to be prevalent On average 34% messages b each newsgroup were addressed to at least one other group, and the average cross-posted message targetied 3.1 ofier newsgroups. Cross-posting was not specific eitiex newsgroups were hig~y inconsistent in tie set of extemrd groups they cross-posted to. Each newsgroup oved cross-posted to a mean of 272 distinct groups. To tier tivestigate the precision of cross-posts, we cdctiated the mean number of cross-posts per distinct cross-posted group. Strong ~ betieen groups wodd lead to a large number of cross-postings to a sd number of groups and hence a high meu However the mean was 5.4. me absence of specfic cross-posting argues against the view tiere are strong communication ties between specific newsgroups. TESTING THE COMMON GROUND MODEL We next use the common ground model to derive a spec~c set of predictions. Each of the conversational strategy variables of (a) increased FAQ production, @) decreased cross-posting; (c) longer messages can be seen as an index of common ground. From our eartier arguments we would therefore expect demographics to affect common ground and hence conversational strategim in the following ways: FAQs were detected autornaticdy by an algorithm which searched subject ~es for tie words “FAQ” or ‘Trequendy Asked Questions”. It excluded certain other special cases such as subject ties including the word “re”. FAQs are a common featnre of newsgroups which is shown by the fact that 54% of newsgroups had FAQs, although less than 0.5?4 of messages were FAQs. We rdso cdctiated average nzessagelength, wtich was 44 hes. Hl~b,c Newsgroups that are larger in size %ti have more Bcdty establishing common ground, we would therefore exTect them to have (a) fewer FAQs; @) more crosspostin~ and (c) shortermessages; H2~b,~ Newsgroups containing many repeat posters (i.e. people who arefamiIiar with the newsgroup) will estabfish common ground more easily, they wi~ therefore have (a) more FAQs; @) less cross-posting; and (c) longer messages; ~60 ——— -——. —— —— Afoderation -—— — -— .->:: _- FAQs 0.16 F 1 i? O*M Figure 3: Causal model sho~vingeffuts of vanabIa in structural quation analysis,for demographic, conversational strate~ and interactivip van-abla H9: Moderated newsgroups * have greater interactivity as dested by increased thread depti, H3~b,c J[oderated newsgroups shodd more easfiy estabfish common grom~ and hence have (a) more FAQs; @) less cross-posting and (c) longer messages; We tested these predictions using tie fo~owing causal mode~ which was tested using a series of regression tiyses [4]: Tt7e fio exTected conversational strateg variables to newsgroups can estabkh tiuence interactivity. E common ground tiough effective communication strategies, we shotid expect this to be Mested in terms of more interactive conversation as indicated by greater tiead depti Demographics + Conversatioti strategies+ kteractivity I 1 me oved anrdysis involved two steps. k the fist we regressed each of the demographic variables onto the conversational strate~ variables to test H1-H3, and second we regressed W variables onto interactivi~ to test H4-H9. Where relevant variables were normafised to allow for the fact that there were different numbers of messages and participants contributing to each news~oup. For several variables (e.g. cross-posting) we experimented with Merent operationatiations (e.g. mean number of crosspostimessage, mean overa~ number of groups crossposted to), choosing the operationrdisation that best accounted for the variance in the regression equations. H* NTewsgroups with more FAQs ~ have greater interactivity as Mested by increased threaddep~ = Newsgroups with less cross-posting ti have greater interactivity as -ested by increased thread depti, H& Newsgroups with longer messages ti have greater titeractivi~ as manifested by increased thread dep~ We shotid dso exTect demographic factors to have direct &ects on interactivity: NTewsgroupsthat are larger in stie ti have less ~ interactivity as manifested by reduced thread dep@ : Effects of demographics on conversational strategy Figure 3 shows the results of three regressions of HS: NTewsgroupscontaining repeat posters (ie. people who with the newsgroup) ~ have greater are fmiliar interactivity as manifested by increased tiead depti, demographics on conversational strategy. me numbers on the arcs represent the standardised beta weights. For b 261 . . . ..- -. . .. ., ..,. I simpEcity of presentation we show ody significant relationships. Ovefi the modek for each of tie three factors was statisticdy significmt For message length (~3,49q = 413, p <0.001, R2 = 020), for cross-posting (f13,49q = 11.3, p <0.001, R2 = O.OQ, and for FAQs (~3,49q = 5.89, p <0.01, R2 = 0.03). We now discuss tie spec%c predictions about the effects of each demographic factor. We stie had no effect on FAQs @la), predictions concerning hypothesis Hlb and Hlc were cofied newsgroups that are hger in she have initiating attempts. This suggests a problem of conversational ineti.a: it seems to be problematic to starta mass interaction but it is rehtively easy to continue it once started We &o found massive pati-c~ation inequaliti~ between ditTerentpeople in a given newsgroup. This is an important clarification of the view that the ktemet is an egatitian fomm [9]. We it is true that anyone can post to a newsgroup, the descriptive data clearly show that conversations in newsgroups are dominated by a minority of hifly verbose participants.Again this argues for a view that Usenet conversations do not stricdy involve repeated levek of mass participatio~ as evidenced by the fact that 27% of messages are generated by people who ody contribute once to the newsgroup. OveraU these retits suggest that as far as active posting is conceme~ mass interaction may be a tinomer. Typical conversations involve ody sma~ numbers of posters, and newsgroups are often dominated by cfiques of verbose contiiutors. This pardcipation inequality contrasts with research on face-toface interaction and video-mediated communication which has shown much more equal levels of participation [14]. Participation inequality may not have whofly negative outcomes, however. The majority of Usenet participants may wefl be satisfied with making tiequent contributions, posting ody about issues that are important to them. Such peripheti participation may enable people to remain in the backgroun~ monitoring general conversations unti they spot a topic of direct relevance, or they need to pose a question of tieir own. This type of background involvement has been observed to be beneficial in the case of corporate ed [11] and interpersonal communication [12]. This form of participation has also been proposed as an important way for novices to learn about a novel topic [12]. more cross-posting (t= 2.2, p <0.05, B= 0.10) and shorter messages (t = 3.2, p c 0.01, P= 0.14). The f~ty predictions H2b and H2c were rdso confirmed newsgroups containing more famZiar contributors had less crossposting (t= 4.5, p <0.001, B= 0.02) and longer messages (t= 2.4, p <0.05, P= 0.01), although again there was no fiect of ftiarity on FAQs, so H2a was not cofie~ Fin~y predictions H3<b, and c were W confirme& Jfoderation led to more FAQs (t= 3.5, p <0.001, O.1~, reduced cross-posting (t = 2.8, p c P= 0.01, ~= 0.13), and longer messages (t = 9-5, p < 0.001, ~= 0.39). Effec& of demograpties and conversational s~tegy on interactivity model was ve~ successti at predicting The ~ titimCtiVi@ (f16,493] = 133.9, P <0.001, R2 = 0.62). However, two of the predictions about the effects of conversation strategy on interactivity were &confirmed H5 predicted that newsgroups with less cross-posting@ have greater interactivity, but in fact tie less cross-posting there w~ in the group, the Ims the threading (t = 23.3, p < 0.001, )= 0.73). H6 predicted thatnewsgroups witi longer messages d have greater intemctivity, but in fact message len@ reducd interactivity (t= 52, p <0.001, &;~S). There were no effects of FAQs on interactivi~ J~e the conversational and participation data suggest a view in which conversation is carried out Iocafly by a few participants, the cross-posting data suggest a s~ghtiy ~erent perspective. It turns out that cross-posting is fiequen~ although it propagates potentia~y irrelevant messages and is thus contrary to both the common ground model and Netiquette guidehes. Cross-posting can also inform us about the relations between newsgroups. When people do cross-pos< they are unspecific about the groups they cross-post to, as indicated by the fact that the average newsgroup cross-posted to over 200 others. This in turn argues against the view that there are tight conversational M between ~erent newsgroups [15]. Ody one of the predictions about the effects of was confirme~ namely H8, demographi~ on interacti~ newsgoups containing f~ar contributors ~ have greaterbead deph (t= 16.2,p <0.001, B= 0.05). Neither FR nor H9 concerning the effect of FAQs and stie were vfidated CONCLUS1ONS We cm draw a number of conclusions about mass interaction from this researck Mthough there are large numbers of people contributing to each newsgroup and large numbers of messages ove~ the descriptive data show that the defadt mass interaction consists of a 43 he message referring to two previous messages. This indicates moderate but not large amounts of interactivity. This geneti view about interactivity shotid be tempere~ however, by the fact that many attempts to initiate interaction were fdures; over 40°A of messages were We *O tested a number of predictions derived from the common ground model. These were matiy verified for the effects of demographics on common ground. The demographic variables of moderation fdarity and der newsgroups au increased common ground as measured by increased message length and decreased 262 : ,- ● : ~ ~ ~ ~ ;“ i. I 1 ! : , ~ 1 ~ ‘ - >. -, !.-. -—.—————- .--., ~ . — . anrdysisof FAQs and moderators’ strategiesis necess~ determinewhether these suggestions are the case. cross-posting. me effects on FAQ production were less clear however. My moderation increased FAQ prOductiO~ with neither ~ty nor she having an effect ~ view argues against the view that FAQs are a “defence mecti~” tiing from participants’ desire to Our &dings shodd ako be qutied by a number of provisos. Our stricfly quantitative anrdysis needs to be complemented by content tiysis, as wefl as surveys of Usenet participants.The cwent dysis is silent about the effects of conversatioti contenti which specific topics or conversationrd styles encourage large responses and which fd to eticit a response? ~t factors encourage or discourage flaming? mat are people’s reactions to flagrant examples of cross-posting? Surveys and interviews could *O addresspeople’s attitudesto, and satisfaction with their Usenet interactions. My do people contribute to certain discussions but not to others? How long do people lurk before they tit post? And how is dominance viewed? my do certain people post mdtiple messages and how are they perceived by others for doing so? Again some of these issues have been addressed in sma~ scale studies, but more of ti me of work is needed [1,17]. There are also issues about moderation and FAQ maintenance which are higMy relevant for issues of group memory [2]: why do certain people take on the responsib%ty of moderating or maintaining FAQs, and what is their motive for doing so? Other issues that need to be addressed by content analysis include deh%erateattemptsto subvert Usenet conversation or the generation of messages by rw amrning, “trohg”) Mcial agents ~bots”). insfl consensus in diverse or rapidy changing user poptiations. Mther FAQs are most ~ely to arise under the orderly conditions induced by moderation Jf%at about tie effects of derno~phics on interactivity? Here the common ground model f~ed much less wek ody f~arity directiy increased heading, with neither moderation nor ske having an effect Furthermore, and contrary to our expectation, interactivity. ~s is counter cross-posting increased to both the common ground view and the NTetiquetteguidehex both of these contend that cross-posting shotid Mute conversatiomd focus and hence reduce interactivity. This retit is consisten~ however, with a ~erent perspective on large scale interactio~ namely }veak ti~ [5,@. The positive effects of aoss-posting on interactivity suggests that people exploit the mass distribution properties ~a diverse poptition of Usenet to go beyond a particdar newsgroup to carry out their conversations. Combining the retits about the effects of f~arity and cross-posting on interactivity indicates thatthere may be ~o complementary sets of circumstances hat facfitate mass titeractiom me first occurs when ftiar participants share common ground and the second is where people seek out diverse perspectives by posting outside their newsgroup. Our M prediction about the effects of message length on interactivity was *O &co*e& shorter messages acdy promoted interactivity. How can we explain this? One possl%fity is consistentwith communication overload [11, 18]. Given tie huge amounts of conversation tic in Usene~ people are less ~ely to rea~ and hence reply to, long messages. Such an interpretation wodd be consistent witi the Netiquette strictore ~avoid long postings”). FMy there are issues concerning the use of thread depth to measure interactivity. One potentird objection to using threading is it f~ to include ‘%ackchannel” responses in emaik Usenet users report that they sometimes reply to a pubfic Usenet message privately in emafl. However, our interesthere was in pubhcly observed mass interaction, and ed conversations are not part of the pubfic record. This argument *O appties to the issue of “lurking”, i.e. reading newsgroup messages without responding to them. me lurking may be a prevalent behaviour, again it leaves no pubfic conversationrdtrace. FinaUy there is the question of “flaming”. Deeply threaded Usenet interactions sometimes retit from emotiona~y charged and occasional persotiy abusive exchanges. However this addresses the question of conversational conten~ an issue which we leave to fntore work Mthough moderation and FAQ production increased common groun~ they ha no effecs on interactivity. my was this tie case? One reason why moderation does not increase threading may be that moderatom incrae conversatioti relevance by deciding that a new posting is tan~enti to tie topic, and stopping the current thread h doing so, however, they automaticdy reduce the amount of threading. A reason for the absence of expected conversatiomd benefit of FAQs may be that FAQs have two contradictory effects. FAQs may promote conversation by providing access to information about group titure, conversatioti expectations ~a a precis of group conversatioti history. At the same time they may discourage newcomers from productively revisiting a previous conversation out of the mistaken be~ef that the topic has been efiausted New research involving content mat are the theoreticrd implications of these restits? Our data show thatw~e the common ground model provides a good account of the effects of demographics on conversational strategy, it is much weaker at predicting interactivity. Two major mo~cations need to be made in applying the common ground model to mass interaction. The fact that shorter messages promoted interactivity suggests that a model of mms interaction also needs to incorporate the notion of conversational overload participants have to flter large numbers of messages to fia relevant inforrnatio~ with the consequence that long 263 —-. to -. ● —-— . .. . . .,-.,- ..__ messages may be ignored Our finding that cross-posting benefits interaction indicates that he model needs to take account of tie benefits of both -arity and diversity. On fie one hand common ground can promote consensus but conversations can grow stie through ovefitiarity of topics and people. On tie other han~ a diverse popdation may stirndate interaction but their widely disparate perspectives may mean tit no conversatioti progress can be made. ACKNOWLEDGMENTS We thank hfarc Smi@ Lyn W~er and Jtia Hirschberg for comments on early versions of this work REFERENCES 1. Ban N. titerpreting soap operas and creating community: bide a computer-mediated fa ctitie, Journal ofFolHore Research, 30,143-176,1993. 2. Berw L., Jefies, R, O~ay, V., Paepcke, A., and J$%arto$ C., were did you put it? ksues in the design and use of a group memory. h Proceedings of CH’93, 33-30,1993. 3. Clark K Arenas of language use. University of ticago Press, Chicago, 1992. 4. Cohe% P. Empirical methods for artiicial intelligence, Bosto~ h~ Press, 1995. 5. Constan$ D., Spro~ L., and Kiesler, S. The kindness of stigers, Organtiational Science, 7, 119-135, 1996. 6. Grano\’etter,hL me strength of weak ties. American Journal ofSociolo~, 7S, 1360-13S0, 1973. 7. Harrisoq hI. me Usenet handbook O’Re~y, 1994. 8. Hortoq hL (1993): Rties news.rnisc,news.answers.,~ Bosto~ for posting to Usene~ at timi~edw 10.Kofloc~ P and Smiti, M. Managing the Virtual Cornmom h S. Herring (Ed.) Computer-Mediated Communication, John Benjamins, Ptiadelphia, 1996. ‘“ 1l. Krau~ R and Attewe~, P. Media use in a global corporation. k S. Kiesler (Ed.), Culture of the Internet, Erlba~ NJ., 1997. 12.Krauq R, Fish R, ROOLB. and ChaKonte, B. Mormal communication in orgtiations. k R Baecker @d.), Groupware and Computer Supported Co-operative Work, 2S7-314, Morgan KaCA, 1992. 13.Marti, M. Towards a critical mass theory of interactive media. h J. Fuk and C. Stetileld (Eds.), Organtiations and Communication Technolog, Sage, CA, 1990. 14. O,Con@ B., Whittaker, S., and WSbur, S. Conversations over video conferences, Human &mputerInteraction, S, 3S942S, 1994. 15.Smi@ M. Measuring the social s~cture of the Usene~ Unpubbhed paper, University of CaMornia, Los Angeles, 1997. 16.Sprod, L., and Kiesler, S. Connections, Boston, MT Press, 1991. ,‘ 17.Suttoq L. Gender, power, and sflencing in electronic discourse on Useneg Proceedings of the 20th Berkel~ Linguistics Socie@, University of California, Berkeley, 1994. lS.~ttaker, S. and Sidner, C. Ed overload exploring personal information management of emafl. h Proceedings of CH96, ACM Press, ~, 1996. 19.Whittaker, S. T*g to strangers: an evaluation of the electronic collaboration, h factors tiecting Proceedings of CSCW96, ACM Press, ~, 1996. I I 9. King, J. Grinter, R and Pickefig, J. The rise and fM of NehNe. h S. Kiesler @&), Culture of the intemet. Erlba~ NJ, 1997. I I 1 I .—. _ —. —